xfs: only remap the written blocks in xfs_reflink_end_cow_extent
PATCH 1/1 xfs: only remap the written blocks in xfs_reflink_end_cow_extent
xfs_reflink_end_cow_extent looks up the COW extent and the data fork extent at offset_fsb, and then proceeds to remap the common subset between the two.
It does however not limit the remapped extent to the passed in [*offset_fsbm end_fsb] range and thus potentially remaps more blocks than the one handled by the current I/O completion. This means that with sufficiently large data and COW extents we could be remapping COW fork mappings that have not been written to, leading to a stale data exposure on a powerfail event.
We use to have a xfs_trim_range to make the remap fit the I/O completion range, but that got (apparently accidentally) removed in commit df2fd88f8ac7 ("xfs: rewrite xfs_reflink_end_cow to use intents").
Note that I've only found this by code inspection, and a test case would probably require very specific delay and error injection.
--- fs/xfs/xfs_reflink.c | 1 + 1 file changed, 1 insertion(+)
fs/xfs/xfs_reflink.c +1 -0
@@ -784,6 +784,7 @@ xfs_reflink_end_cow_extent( } } del = got; +xfs_trim_extent (&del, *offset_fsb, end_fsb - *offset_fsb);/* Grab the corresponding mapping in the data fork. */ nmaps =1 ;
Signature · 1 line
2.39.2
Quoted 11 lines · Christoph Hellwig
xfs_reflink_end_cow_extentlooks up the COW extent and the data fork extent atoffset_fsb, and then proceeds to remap the common subset between the two.It does however not limit the remapped extent to the passed in [
*offset_fsbmend_fsb] range and thus potentially remaps more blocks than the one handled by the current I/O completion. This means that with sufficiently large data and COW extents we could be remapping COW fork mappings that have not been written to, leading to a stale data exposure on a powerfail event.
Quoted 6 lines
We use to have a
xfs_trim_rangeto make the remap fit the I/O completion range, but that got (apparently accidentally) removed in commitdf2fd88f8ac7("xfs: rewritexfs_reflink_end_cowto use intents").Note that I've only found this by code inspection, and a test case would probably require very specific delay and error injection.
Hmm. xfs_prepare_ioend converts unwritten cowfork extents to written prior to submit_bio. So I guess you'd have to trick writeback into issuing totally separate bios for the single mapping. Then you'd have to delay the bio for the higher offset part of the mapping while allowing the bio for the lower part to complete, at which point it would convey the entire mapping to the data fork. Then you'd have to convince the kernel to reread the contents from disk. I think that would be hard since the folios for the incomplete writeback are still uptodate and marked for writeback. directio will block trying to flush and invalidate the cache, and buffered io will read the pagecache.
So I don't think there's an actual exposure vector here, but others can chime in about whatever I've missed. ;)
Quoted 2 lines
Fixes df2fd88f8ac7 ("xfs: rewrite xfs_reflink_end_cow to use intents")Signed-off-by Christoph Hellwig <[email protected]>
Yep, that was a goof, thanks for catching that.
--D
Quoted 19 lines
--- fs/xfs/xfs_reflink.c | 1 + 1 file changed, 1 insertion(+)fs/xfs/xfs_reflink.c +1
-0@@ -784,6 +784,7 @@ xfs_reflink_end_cow_extent( } } del = got; +xfs_trim_extent (&del, *offset_fsb, end_fsb - *offset_fsb);/* Grab the corresponding mapping in the data fork. */ nmaps =1 ;Signature · 1 line
2.39.2
Quoted 3 lines · Darrick J. Wong
Hmm.
xfs_prepare_ioendconverts unwritten cowfork extents to written prior tosubmit_bio. So I guess you'd have to trick writeback into issuing totally separate bios for the single mapping.
Yes. Hitting IOEND_BATCH_SIZE seems like the least difficult one to hit, but even that would require work.
Quoted 4 lines
Then you'd have
to delay the bio for the higher offset part of the mapping while allowing the bio for the lower part to complete, at which point it would convey the entire mapping to the data fork.
Shouldn't really matter which side is faster.
Quoted 5 lines
Then you'd have to convince
the kernel to reread the contents from disk. I think that would be hard since the folios for the incomplete writeback are still uptodate and marked for writeback. directio will block trying to flush and invalidate the cache, and buffered io will read the pagecache.
I don't think on a live kernel it is possible. But if one of the two bios completed before the other one, and power failed just inbetween.
Quoted 15 lines · Christoph Hellwig
Quoted 3 lines · Darrick J. Wong
Hmm.
xfs_prepare_ioendconverts unwritten cowfork extents to written prior tosubmit_bio. So I guess you'd have to trick writeback into issuing totally separate bios for the single mapping.Yes. Hitting
IOEND_BATCH_SIZEseems like the least difficult one to hit, but even that would require work.Quoted 4 lines
Then you'd have
to delay the bio for the higher offset part of the mapping while allowing the bio for the lower part to complete, at which point it would convey the entire mapping to the data fork.Shouldn't really matter which side is faster.
Quoted 8 lines
Quoted 5 lines
Then you'd have to convince
the kernel to reread the contents from disk. I think that would be hard since the folios for the incomplete writeback are still uptodate and marked for writeback. directio will block trying to flush and invalidate the cache, and buffered io will read the pagecache.I don't think on a live kernel it is possible. But if one of the two bios completed before the other one, and power failed just inbetween.
Ooooh, yeah. That could happen if the ioend metadata update gets written to the log device and the system crashes before that other bio even gets a chance to execute.
--D